DocsOverview
Experiment TrackingOverview

Experiment Tracking

Polyaxon tracking records what happens during a run and makes it available in the UI, CLI, and API. Use it to log metrics, metadata, artifacts, custom charts, and lineage from local code, notebooks, and in-cluster jobs.

Log metrics, parameters, and artifacts from any training script with the Polyaxon Python SDK.

from polyaxon import tracking

tracking.init()
tracking.log_inputs(learning_rate=0.001, batch_size=64)

for epoch in range(100):
    loss, acc = train_epoch(model, dataloader)
    tracking.log_metrics(step=epoch, loss=loss, accuracy=acc)

tracking.log_model_ref(model, name="classifier")

Compare experiments side-by-side — filter by metrics, parameters, tags, or status using the query language.

# List top runs by accuracy
polyaxon ops ls -q "metrics.accuracy: >0.95, status: succeeded" \
  --sort="-metrics.accuracy" --limit=10

Built-in support for TensorBoard, Plotly, and custom visualization plugins. Launch TensorBoard as a service alongside your training jobs.

# polyaxonfile.yaml
version: 1.1
kind: operation
component:
  run:
    kind: service
    ports: [6006]
    container:
      image: tensorflow/tensorflow:latest
      command: [tensorboard, --logdir=/outputs, --bind_all]

Track and version any output artifact — models, datasets, plots, and logs. Download artifacts via CLI or SDK.

# Download artifacts from a run
polyaxon ops artifacts --path ./outputs/

# List artifacts for a specific run
polyaxon ops artifacts -uid <run-uuid>

Logging

Start with logging to instrument Python code with the tracking API. For API-level details, see the tracking client and tracking module references.

Visualizations

Use visualizations for native charts, custom plots, and programmatic run comparison.

Runs dashboard

Use the runs dashboard to inspect run info, logs, artifacts, lineage, comparisons, dashboards, timelines, and monitoring views.